Page 15 of 18
Above Fig.: AdvProp improves image recognition. By training model on ImageNet, AdvProp helps
EfficientNet-B7 to achieve 85.2% accuracy on ImageNet, 52.9% mCE (mean corruption error, lower is
better) on ImageNet-C, 44.7% accuracy on ImageNet-A and 26.6% accuracy on Stylized-ImageNet,
beating its vanilla counterpart by 0.7%, 6.5%, 7.0% and 4.8%, respectively. Theses sample images are
randomly selected from category “goldfinch.”
In this paper, rather than focusing on defending against adversarial examples, we shift our attention to
leveraging adversarial examples to improve accuracy. Previous works show that training with adversarial
examples can enhance model generalization but are restricted to certain situations—the improvement is
only observed either on small datasets (e.g., MNIST) in the fully-supervised setting [5], or on larger
datasets but in the semi-supervised setting [21, 22]. Meanwhile, recent works [15, 13, 31] also suggest
that training with adversarial examples on large datasets, e.g., ImageNet [23], with supervised learning
results in performance degradation on clean images. To summarize, it remains an open question of how
adversarial examples can be used effectively to help vision models.
We observe all previous methods jointly train over clean images and adversarial examples without
distinction, even though they should be drawn from different underlying distributions. We hypothesize
this distribution mismatch between fresh examples and adversarial examples is a key factor that causes
performance degradation in previous works.
Q8. Advancing NLP with Cognitive Language Processing Signals
Answer:
When reading, humans process language “automatically” without reflecting on each step — Humans
string words together into sentences, understand the meaning of spoken and written ideas, and process
language without overthinking about how the underlying cognitive process happens. This process
generates cognitive signals that could potentially facilitate natural language processing tasks.
In recent years, collecting these signals has become increasingly accessible and less
expensive Papoutsaki et al. (2016); as a result, using cognitive features to improve NLP tasks has become
more popular. For example, researchers have proposed a range of work that uses eye-tracking or gaze
signals to improve part-of-speech tagging (Barrett et al., 2016), sentiment analysis (Mishra et al., 2017),
named entity recognition Hollenstein and Zhang (2019), among other tasks. Moreover, these signals have
been used successfully to regularize attention in neural networks for NLP Barrett et al. (2018).
However, most previous work leverages only eye-tracking data, presumably because it is the most
accessible form of cognitive language processing signal. Also, most state-of-the-artwork(SOTA) focused
on improving a single task with a single type of cognitive signal. But can cognitive processing signals
bring consistent improvements across modality (e.g., eye-tracking and EEG) and across various NLP